Potato crop stress early warning from thermal anomalies
Thermal infrared sensors detect canopy cooling and warming anomalies in potato fields days before visible symptoms appear. Combined with ERA5 humidity reanalysis, field-level risk scores give agronomists time to act.
Sensors
- Landsat 8/9 TIRS: Two thermal infrared bands centred near 10.9 µm and 12.0 µm; 100 m native resolution resampled to 30 m in data products. Revisit 16 days per satellite, 8 days combined. Daytime overpass only (approximately 10:00 local solar time). Primary source for field-scale land surface temperature (LST) retrieval.
- ECOSTRESS (ISS-mounted): Five thermal infrared channels from 8.28 to 12.13 µm; 38 × 69 m pixel footprint. Non-sun-synchronous orbit produces variable overpass times including night, enabling diurnal temperature cycle sampling. Revisit irregular, roughly 3–5 days at mid-latitudes. Designed specifically for plant water-stress detection via evapotranspiration.
- Sentinel-3 SLSTR: Dual-view thermal channels at 10.85 µm and 12.0 µm; 1 km spatial resolution in thermal bands. Daily global revisit. Best suited to regional stress indexing and anomaly context rather than individual field diagnosis below roughly 50 ha.
- ERA5 reanalysis (ECMWF): Hourly gridded atmospheric fields at approximately 31 km resolution: 2 m dewpoint temperature, relative humidity, surface pressure and wind. Used to compute vapour pressure deficit (VPD) and infection-period humidity thresholds that contextualise thermal anomalies.
Why temperature tells you something before the lesions do
A healthy potato canopy transpires actively, keeping leaf surface temperature 2–5 °C below air temperature on a clear day. When Phytophthora infestans begins colonising tissue, stomatal function is disrupted before necrosis is visible. The canopy warms relative to healthy neighbours. Conversely, free moisture on leaves from high-humidity nights can suppress transpiration and produce anomalous cooling that, combined with the right dew-point conditions, signals an infection-period has occurred. Either signature, warming or anomalous cooling, can precede visible blight by three to seven days under published field-trial conditions.
That window is exactly what thermal remote sensing is designed to catch. The physics are straightforward: emitted radiance in the 10–12 µm atmospheric window is a direct function of surface temperature, and retrieval algorithms (split-window or single-channel, depending on the sensor) convert at-sensor radiance to land surface temperature with uncertainties typically cited at 1–2 °C for Landsat TIRS and better than 1.5 °C for ECOSTRESS under clear skies. One to two degrees matters here: the stress signal is real but not large, which is why honest uncertainty accounting is non-negotiable.
What each sensor actually delivers, and where each one fails
Landsat 8 and 9 TIRS offer the longest consistent thermal archive (Landsat 8 launched 2013) and the finest spatial resolution in freely available thermal data at 100 m native. For potato fields larger than roughly 5–10 ha, a single-field LST value is retrievable. Below that, mixed-pixel contamination from field margins, access tracks and bare soil between rows degrades the signal. The 16-day revisit per satellite is the more serious constraint: a disease-permissive weather window can open and close in 48 hours, and Landsat may simply not be overhead.
ECOSTRESS partially addresses revisit by virtue of the ISS orbital precession, but 'irregular' is the honest description. Data gaps of a week or more are common at any given location, and the ISS orbit does not cover latitudes above approximately 51.6° N or S. For high-latitude potato-growing regions, ECOSTRESS is unavailable. Its finer footprint (38 × 69 m) is genuinely useful for resolving within-field variability in larger paddocks.
Sentinel-3 SLSTR provides daily coverage but at 1 km thermal resolution. A 1 km pixel in a fragmented potato-growing landscape contains multiple fields, multiple crop types and potentially roads and buildings. It is useful for regional anomaly mapping, flagging districts where conditions are deteriorating, but it cannot diagnose an individual field. Cloud is the shared enemy of all thermal sensors: optical depth in the thermal infrared is not dramatically better than in the visible, and a persistent overcast during a high-risk period can black out all three sensors simultaneously. Synthetic aperture radar does not carry thermal channels, so there is no cloud-penetrating substitute for LST.
Turning humidity reanalysis into an infection-period layer
The Beaumont and Smith-Hirst criteria, published in the mid-twentieth century and still used operationally, define a late-blight infection period as any interval during which temperature stays between 10 °C and 24 °C and relative humidity exceeds 75% for at least 11 consecutive hours. ERA5 provides the hourly gridded humidity and temperature fields needed to compute these thresholds continuously, at no cost, with a latency of roughly five days for the standard product and under 24 hours for the ERA5-Land preliminary release.
The practical workflow is to run the Beaumont criterion across ERA5 cells covering the potato-growing area, flag any field that has experienced one or more infection periods in the preceding 72 hours, and then cross-reference that flag with the most recent cloud-free LST anomaly from TIRS or ECOSTRESS. Fields that score positive on both layers, humidity-driven infection risk and a thermal departure from the field's own seasonal baseline, receive the highest alert tier. Fields with infection-period humidity but no thermal anomaly yet are flagged as watch status. The combination reduces false positives from either input alone.
The resolution floor and what it means in practice
No freely available thermal sensor resolves individual potato rows. At 100 m (Landsat TIRS), a pixel covers roughly one hectare. A disease outbreak beginning in a corner of a 3 ha field may occupy less than a third of a single pixel for several days before the thermal signal is detectable above retrieval noise. This is not a solvable problem with current open-data constellations. It is a reason to treat satellite thermal data as a field-level triage tool rather than a within-field diagnostic.
Commercial thermal sensors such as those on some Planet SkySat or Airbus Pléiades Neo configurations do not carry thermal infrared channels. Drone-mounted thermal cameras can resolve row-level stress at centimetre scale, but they require deployment decisions. The satellite layer's role is to tell an agronomist which fields are worth deploying a drone to, or which fields warrant an immediate ground inspection, before visible symptoms have declared themselves across the canopy.
Compositing a risk score and delivering it at field scale
A workable field-level risk score combines four inputs: the z-score of the current LST relative to a rolling 30-day baseline for the same field (normalising for crop growth stage and local climate), the ERA5 infection-period count in the preceding 72 hours, the vapour pressure deficit at the time of the satellite overpass (low VPD amplifies the thermal stress signal's ambiguity), and a crop-stage weight derived from planting-date priors. The output is a dimensionless score, typically binned into three alert tiers, delivered as a GIS polygon layer attributed per field boundary.
Field boundaries are a prerequisite. Where cadastral data are unavailable, boundaries can be derived from multispectral time-series segmentation, but that is a separate analytic step covered in the sibling page on smallholder field boundary delineation. Satellize runs this thermal anomaly pipeline on open Landsat and ECOSTRESS archives and has applied analogous crop-stress scoring methods in its Tonga crop-estimation programme, where field-level data sparsity presents similar boundary and resolution challenges.
Latency from satellite overpass to delivered risk score is typically 24–48 hours for Landsat (after USGS processing) and variable for ECOSTRESS depending on ISS downlink scheduling. ERA5 preliminary fields arrive within 24 hours of the analysis time. An operational system should be designed around these latencies: a 48-hour-old thermal observation is still actionable if the agronomist can reach the field within the disease incubation window.
Typical figures
| Best available thermal spatial resolution | 100 m native (Landsat 8/9 TIRS); 38 × 69 m (ECOSTRESS); 1 km (Sentinel-3 SLSTR) |
| Revisit frequency | 8 days combined Landsat 8+9; irregular ~3–5 days ECOSTRESS; daily Sentinel-3 SLSTR |
| LST retrieval uncertainty (clear sky) | ±1–2 °C Landsat TIRS; <1.5 °C ECOSTRESS (published mission specs) |
| Thermal spectral bands | 10.6–11.2 µm and 11.5–12.5 µm (Landsat); 8.28–12.13 µm five channels (ECOSTRESS); 10.85 µm and 12.0 µm (SLSTR) |
| Cloud penetration | None. All thermal infrared sensors are blocked by cloud cover. |
| Humidity reanalysis resolution | ~31 km grid, hourly (ERA5); latency ~5 days standard, <24 h preliminary |
| Minimum detectable field size (thermal) | ~5–10 ha for a clean Landsat TIRS pixel; ~0.3 ha for ECOSTRESS under ideal geometry |
| Archive depth | Landsat TIRS from 2013 (L8) and 2021 (L9); ECOSTRESS from 2018; ERA5 from 1940 |
| Data access | Free: USGS EarthExplorer (Landsat), NASA Earthdata (ECOSTRESS), Copernicus Dataspace (Sentinel-3), ECMWF CDS (ERA5) |
| Delivered formats | GeoTIFF LST anomaly rasters; GeoJSON/Shapefile field-polygon risk scores; CSV alert tables; optional WMS tile feed |
Analytics Satellize can run
| Field-level LST anomaly map | Split-window or single-channel LST retrieval; z-score against 30-day rolling baseline per field polygon | GeoTIFF and attributed polygon layer, updated per cloud-free overpass |
| ERA5 infection-period count layer | Beaumont criterion applied to hourly ERA5 dewpoint and temperature grids; count of qualifying 11-hour windows in preceding 72 h | Gridded raster and field-polygon summary, daily update |
| Composite blight risk score | Weighted combination of LST z-score, infection-period count, vapour pressure deficit and crop-stage prior; binned to three alert tiers | Field-polygon GIS layer with numeric score and alert tier; CSV for integration into farm management systems |
| Seasonal thermal baseline | Multi-year Landsat TIRS time-series compositing per field; percentile-based normal envelope for each calendar week | Per-field baseline statistics table; used as denominator for in-season anomaly scoring |
| Cloud-gap interpolated LST estimate | Temporal interpolation between bracketing clear-sky observations; flagged with uncertainty band proportional to gap length | Gap-filled raster with per-pixel confidence flag; explicitly labelled as modelled, not observed |
| Weekly agronomist alert bulletin | Automated digest of highest-scoring fields from composite risk layer, ranked by score and area | PDF or structured email report listing field IDs, coordinates, risk tier and recommended ground-check priority |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.